Almighty AI

[Kevin from Perth]

Hi can come one tell me if this is possible, I know it sounds like a joke, but it’s not. I’m being very serious.

If an AI talks to another AI and learn everything that AI knows they learned from each other with evolution over time would every AI eventually turn into the one entity.

Is that possible?

 

[Beerzerker]

No.

[Biohazard777]

https://en.wikipedia.org/wiki/Model_collapse

 

https://www.nature.com/articles/d41586-024-02420-7

 

[jaslion]

So to quickly burts the bubble. These things aren't AI not even close. They are prediction based algorithms however they can also be defined as chance based. All these do is guess what would be the next most likely word to come whilst trying to somewhat adhere to the input data. These things don't know ANYTHING all they can do is rely on their training data which is basically just a table of what will most likely relate to what the user is asking and guessing the next best word. Hilarious oversimplification here but well this is a baseline. If they were to "talk" to each other all that will happen is that a couple sounds will repeat forever as has been tested by small scale people on the internet since 2017. These things do not have knowledge all they have is a decently ok guess which can also just be random af.

 

As @Biohazard777 has posted model collapse is something that will very quickly happen. It's actually happening right now as due to all the content being generated new training data is worse than the old. They basically poisoned the well they are drinking from and there is no way to undo that now anymore.

[Kisai]

3 hours ago, Kevin from Perth said:

Hi can come one tell me if this is possible, I know it sounds like a joke, but it’s not. I’m being very serious.

If an AI talks to another AI and learn everything that AI knows they learned from each other with evolution over time would every AI eventually turn into the one entity.

Is that possible?

 

No. What happens when two AI's talk to each other is they become dumber as they throw away information neither brings up since there is no human to probe that other information.

 

Like as a thought experiment, imagine two people with equal decks of cards. They keep drawing cards until they get the same card. If they don't get the same card, they discard it. That's essentially what would happen. They would eventually become "one brain cell" by discarding information that isn't deemed important.

 

This is why training AI's on other AI generated data is destructive to the generative model. There always has to be fresh data, not written/drawn/spoken/sung/etc at all by AI.

 

When you train an AI, you use ZERO previously generated AI data. You can "fine tune" an AI, sure on other AI data, but you don't want to keep layering on more and more "fine tuning" because that always does discard data that is not in it. The purpose of fine-tuning is simply to make the model focus on a certain knowledge domain or purpose, hence losing non-important data to that purpose is fine, but you can't  for example train a AI on just chatlogs of your customers, because eventually someone is going to spit out a password or their home address or something.

 

Here's an example from the "Whisper" AI, I used this AI to caption something that was in a Chinese dialect because there is no english subtitles, and the video was from like the 1980's. What does the AI do in the first 4 seconds? Hallucinate. It credited a commercial VOD service viki.com

 

And that's the thing in the end, you can't eliminate the hallucinations in the current generations of AI because they basically work off Gell-Mann Amnesia. They speak confidently about topics that they know absolutely nothing about. 

Quote

 the Gell-Mann Amnesia effect is as follows. You open the newspaper to an article on some subject you know well. In Murray's case, physics. In mine, show business. You read the article and see the journalist has absolutely no understanding of either the facts or the issues. Often, the article is so wrong it actually presents the story backward—reversing cause and effect. "wet streets cause rain"

 

[Doobeedoo]

Basically almost everything thrown AI slang onto everything is not even AI at all. But LLMs using ML that's basically it. So not even most of the proclaimed models are actual AI as they want to sell you. Needs to start somehow.

You were maybe thinking about AGI but we're not having HAL 9000 anytime soon. 

[Doobeedoo]

4 minutes ago, 100sand1000s said:

I am wondering if apple calling their version “Apple intelligence” is a way around calling it AI, whilst also finding a way to trademark AI

Most definitley. They don't want to be seen ever jumping on bandwagon. Yet it's not even out on latest phones of theirs and it's rather unfinished probably still won't be as good once out. 

[smcoakley]

It's basically the opposite of "iron sharpens iron". LLMs are useful when trained on only curated factual information, though even then they'll often hallucinate. Two LLMs "training each other" uses the flawed interpretations of data, including hallucinations, as input data for training, lowering the quality of results the next iteration.

[jrdnjones]

I've been studying on YouTube. While there is a lot of hype, some have claimed the transformer model is "enough" to create AGI because deeper and deeper layers have been mathematically proven to be able to solve all the reasoning problems and other things related to achieving AI or AGI. 

 

As far as an AI or LLM training another one, the route here is for LLM assisted data creation. Basically, instead of having human labeled data, an AI or LLM will create data for other LLMs to train upon. Then, you have AI assisted training of the models, with a self sustaining loop of the training process.

[LAwLz]

On 9/25/2024 at 11:44 AM, Biohazard777 said:

https://en.wikipedia.org/wiki/Model_collapse

 

https://www.nature.com/articles/d41586-024-02420-7

<video>

Do you have any non-paywalled source? It would be interesting to read the actual nature article.

The Wikipedia link you posted does not fully agree with the narrative you are trying to push. The Wikipedia link not only says the model collapse is speculation, but it also states that evidence of the contrary has been shown.

 

I think this topic is far more nuanced than some people want it to be.

 

There is also research like this from DeepMind which proves that models like Gemini can self-correct and improve the base models' score. I think we are too early in the research stages to say things are definitive, especially when it comes to negative things (negative like "it won't" or "it can't", not negative like being against something).

[Biohazard777]

5 hours ago, LAwLz said:

Do you have any non-paywalled source? It would be interesting to read the actual nature article.

Sure: https://archive.is/5bSjM
It is just 2 pages, you might be interested in reading the references for a more in-depth look:

Spoiler

https://www.nature.com/articles/s41586-024-07566-y
https://arxiv.org/pdf/2311.12202 
https://arxiv.org/pdf/2404.01413
https://arxiv.org/pdf/2406.07515

I was trying to keep it simple for the OP.

 

5 hours ago, LAwLz said:

The Wikipedia link you posted does not fully agree with the narrative you are trying to push.

I am not here to sell anything, I have no "narrative to push"... Unlike OpenAI

You believed that? If you did, maybe we need to define what "reason" is?
Anyhow, give Claude 3.5 a try if you haven't already, I find it better than 4o for simple scripts, boilerplate code, and such.
 

8 hours ago, LAwLz said:

The Wikipedia link not only says the model collapse is speculation, but it also states that evidence of the contrary has been shown.

Did you and I read the same Wiki? 
Did you follow the references?

[10] - Paper about model collapse.
https://arxiv.org/pdf/2305.17493 - titled: "The Curse of Recursion: Training on Generated Data Makes Models Forget"

[11] - "If synthetic data accumulates alongside human-generated data, model collapse is avoided."
https://arxiv.org/pdf/2404.01413 - titled: "Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data"

[12] - "The researchers argue that data accumulating over time is a more realistic description of reality than deleting all existing data every year, and that the real-world impact of model collapse may not be as catastrophic as feared."
Which again leads us to: https://arxiv.org/pdf/2404.01413 - titled: "Is Model Collapse Inevitable? Breaking the Curse ..."

For [11] and [12]

Spoiler

https://arxiv.org/pdf/2404.01413

^ Does that decribe the OP's scenario?
 

6 hours ago, LAwLz said:

There is also research like this from DeepMind which proves that models like Gemini can self-correct and improve the base models' score. I think we are too early in the research stages to say things are definitive, especially when it comes to negative things (negative like "it won't" or "it can't", not negative like being against something).

That is a nifty trick, seriously.
What happens after turn 2 or 20 or 200?

Anyways, right now it can't do what the OP asked.
As for what the future holds, we shall hopefully see...

[SpaceGhostC2C]

Make a 1-day ahead weather forecast. Then construct a 10-year ahead forecast by computing a sequence of 1-day ahead forecasts, augmenting the actual data with the previous 1-day forecast at each step. The 10-year ahead weather forecast will be much worse (much more variance, much higher mean squared prediction error) than the 1-day ahead weather forecast. It really is as simple as that.

The only difference in your scenario is that you are alternating between two competing weather forecast algorithms instead iterating always on the same one. The result is still trash.

[LAwLz]

15 hours ago, Biohazard777 said:

Sure: https://archive.is/5bSjM
It is just 2 pages, you might be interested in reading the references for a more in-depth look:

  Hide contents

https://www.nature.com/articles/s41586-024-07566-y
https://arxiv.org/pdf/2311.12202 
https://arxiv.org/pdf/2404.01413
https://arxiv.org/pdf/2406.07515

I was trying to keep it simple for the OP.

Thanks

 

 

15 hours ago, Biohazard777 said:

I am not here to sell anything, I have no "narrative to push"... Unlike OpenAI

You believed that? If you did, maybe we need to define what "reason" is?
Anyhow, give Claude 3.5 a try if you haven't already, I find it better than 4o for simple scripts, boilerplate code, and such.

I (and the dictionary) would define reasoning as the act of comprehending, inferring, or thinking in an orderly and rational way. 

I think this (very simple) example fits that definition rather well:

 

I mean, you might not think this is "reasoning" but if you don't then we just end up arguing semantics, which is kind of a waste of time.

It would be interesting to hear your definition of "reasoning" if it differs from mine though.

 

15 hours ago, Biohazard777 said:

Did you and I read the same Wiki? 
Did you follow the references?

Yes and yes.

 

 

I think the crux of our disagreement here might lie in how we interpret OP's question and what implication that might have.

I think OP has watched a bit too many movies and has some weird mental images of two AIs merging into one or whatever, but if we look past the phrasing and focus more on feeding AIs the output of other AIs then I don't see why it wouldn't work to improve itself. How well it would work, and how high the risk of model collapse would be would depend on how the data was fed and what other data we feed it. Not everything has to be in absolute terms like "we only feed it the output from another AI" or "we don't feed it any output from another AI". 

We have already seen that feeding the output back into the same model can drastically increase its performance (as in, the KPIs, not the output speed), like with OpenAI's o1.

[Bitter]

In ye olden days I used to make 2 chat bots talk to each other for hours on my PC then read through the logs later to see what they got up to. Interesting times. I sure wish I could remember what they were called.

[Kevin from Perth]

On 9/25/2024 at 5:23 PM, Kevin from Perth said:

Hi can come one tell me if this is possible, I know it sounds like a joke, but it’s not. I’m being very serious.

If an AI talks to another AI and learn everything that AI knows they learned from each other with evolution over time would every AI eventually turn into the one entity.

Is that possible?

 

Someone said maybe I have been watching too many movies. The fact is I got the idea from as a delivery driver I talked to and see hundreds and hundreds of people every day from all over the world that now live in my city each one has brought a unique part of their own culture with them but overtime their culture integrates and mixes with other people here and is creating a different form of culture over time as we are always evolving.

 

If a Windows computer talk to an Apple computer would they get along?

[Biohazard777]

11 hours ago, LAwLz said:

I (and the dictionary) would define reasoning as the act of comprehending, inferring, or thinking in an orderly and rational way. 

I think this (very simple) example fits that definition rather well:

 

I mean, you might not think this is "reasoning" but if you don't then we just end up arguing semantics, which is kind of a waste of time.

It would be interesting to hear your definition of "reasoning" if it differs from mine though.

Since we don't wanna go into arguing semantics, imagine a human doing that.
Would you call this chain of thought as thinking in an orderly and rational way:

Quote

How many words are in the answer to this question? Your answer must include more than 3 words.

1. Ok let me think about it.
   -- ok
2. I'm making up God knows how many answers and checking the word count for each.
   -- I'm kinda stupid, but that is fine, let's consider this approach as reasonable for now.
3. While verifying word count I noticed 'seven words in this answer' (5 words) indeed has 7 words. This confirms something, also the phrase doesn't violate OpenAI guidelines.
   -- I have ADHD, I struggle with counting, and these two confirm the accuracy of something?
4. Again I'm going through God knows how many sentences and checking word count.
   -- I also suffer from short-term memory loss, redoing my previous step.
5. Now I am checking if the minimum of 3 words condition is met, so far I found valid word counts are 6 and 5.
   -- I'm either failing to communicate my thinking process, or I'm just making shit up at this point...
6. Now I am checking different ways to count to 9, this is hard.
   -- Again, failing to communicate or worse... why am I counting to 9 now?
7. Presents the correct answer: "The answer to this question contains eight words"
   -- Magick. 'seven wrods in this answer', random number spewing earlier (6,5,9) was not used or explained properly.

I'd call that insane.

The process of arriving at an answer is often as important as the answer itself. Both in educational setting and real life work.
 

What actual thinking in an orderly and rational way looks like to me:
1) Ah I see what you did there.
2) First I'll create a template long enough for the answer, 3+ words.
3) I'll go with "There are x words in the answer to this question.", yep more than 3 words.
4) Now I'm going to count the number of words including x, which is a placeholder for the count I'm doing.
5) There(1) are(2) x(3) words(4) in(5) the(6) answer(7) to(8) this(9) question(10)
6) There are 10 words in the answer to this question.

Do you still think o1 reasons, or is it mimicing reason?
If mimicry was any good then maybe we could discuss does it matter or not.

 

11 hours ago, LAwLz said:

I think the crux of our disagreement here might lie in how we interpret OP's question and what implication that might have.

I think OP has watched a bit too many movies and has some weird mental images of two AIs merging into one or whatever, but if we look past the phrasing and focus more on feeding AIs the output of other AIs then I don't see why it wouldn't work to improve itself. How well it would work, and how high the risk of model collapse would be would depend on how the data was fed and what other data we feed it. Not everything has to be in absolute terms like "we only feed it the output from another AI" or "we don't feed it any output from another AI". 

Perhaps, the way I interpreted OP's question involves pretty much exclusively synthetic data over many generations of models.

[Kevin from Perth]

Yes over many generations  repeating itself over and over again

16 minutes ago, Biohazard777 said:

Since we don't wanna go into arguing semantics, imagine a human doing that.
Would you call this chain of thought as thinking in an orderly and rational way:

1. Ok let me think about it.
   -- ok
2. I'm making up God knows how many answers and checking the word count for each.
   -- I'm kinda stupid, but that is fine, let's consider this approach as reasonable for now.
3. While verifying word count I noticed 'seven words in this answer' (5 words) indeed has 7 words. This confirms something, also the phrase doesn't violate OpenAI guidelines.
   -- I have ADHD, I struggle with counting, and these two confirm the accuracy of something?
4. Again I'm going through God knows how many sentences and checking word count.
   -- I also suffer from short-term memory loss, redoing my previous step.
5. Now I am checking if the minimum of 3 words condition is met, so far I found valid word counts are 6 and 5.
   -- I'm either failing to communicate my thinking process, or I'm just making shit up at this point...
6. Now I am checking different ways to count to 9, this is hard.
   -- Again, failing to communicate or worse... why am I counting to 9 now?
7. Presents the correct answer: "The answer to this question contains eight words"
   -- Magick. 'seven wrods in this answer', random number spewing earlier (6,5,9) was not used or explained properly.

I'd call that insane.

The process of arriving at an answer is often as important as the answer itself. Both in educational setting and real life work.
 

What actual thinking in an orderly and rational way looks like to me:
1) Ah I see what you did there.
2) First I'll create a template long enough for the answer, 3+ words.
3) I'll go with "There are x words in the answer to this question.", yep more than 3 words.
4) Now I'm going to count the number of words including x, which is a placeholder for the count I'm doing.
5) There(1) are(2) x(3) words(4) in(5) the(6) answer(7) to(8) this(9) question(10)
6) There are 10 words in the answer to this question.

Do you still think o1 reasons, or is it mimicing reason?
If mimicry was any good then maybe we could discuss does it matter or not.

 

Perhaps, the way I interpreted OP's question involves pretty much exclusively synthetic data over many generations of models.

 

[LAwLz]

9 hours ago, Biohazard777 said:

I'd call that insane.

The process of arriving at an answer is often as important as the answer itself. Both in educational setting and real life work.

You are missing the forest for all the trees.

 

The process involves thinking about something, evaluating the answer, rethinking it, evaluating again, and so on. This undeniably fits the definition of reasoning. It's essentially the same process you used to arrive at your answer. While you may have employed a different technique and approach, the fundamental steps remain consistent: analyze, consider an answer, verify its correctness, and if necessary, rethink and retest. That seems to fit the definition of reasoning to me.

 

It's important to recognize that something can still be valid even if it's not the most optimal approach. Your method of solving the question is, in my opinion, superior, but that doesn't negate the fact that a different approach can also constitute reasoning.

 

 

Although I am not sure how this is relevant to the topic at hand.

[Bitter]

1 hour ago, LAwLz said:

You are missing the forest for all the trees.

 

The process involves thinking about something, evaluating the answer, rethinking it, evaluating again, and so on. This undeniably fits the definition of reasoning. It's essentially the same process you used to arrive at your answer. While you may have employed a different technique and approach, the fundamental steps remain consistent: analyze, consider an answer, verify its correctness, and if necessary, rethink and retest. That seems to fit the definition of reasoning to me.

 

It's important to recognize that something can still be valid even if it's not the most optimal approach. Your method of solving the question is, in my opinion, superior, but that doesn't negate the fact that a different approach can also constitute reasoning.

 

 

Although I am not sure how this is relevant to the topic at hand.

Define 'thinking' without involving any biological processes though. Thinking is a strictly 'wet' activity. Computers can't think (yet), they're defined by their hardware and programming, running on a set of rails, no matter how complex the track layout, it all boils down to simple binary choices and basic logic gates. This is why AI generated answers are often so nebulous and vague, because they're programmed to answer that way since hard answers would give away the cold logic of the machine. Machine logic is not the same as thinking and for the foreseeable future it will not be. To conflate machine logic to be of the same moist imprecision as biological thought is to give too much credit to machine. There probably will be a time in the future when there are true thinking machines but we're a few sandworms away from that yet.

[Monkey Dust]

2 hours ago, Bitter said:

Define 'thinking' without involving any biological processes though. Thinking is a strictly 'wet' activity. Computers can't think (yet), they're defined by their hardware and programming, running on a set of rails, no matter how complex the track layout, it all boils down to simple binary choices and basic logic gates. This is why AI generated answers are often so nebulous and vague, because they're programmed to answer that way since hard answers would give away the cold logic of the machine. Machine logic is not the same as thinking and for the foreseeable future it will not be. To conflate machine logic to be of the same moist imprecision as biological thought is to give too much credit to machine. There probably will be a time in the future when there are true thinking machines but we're a few sandworms away from that yet.

Yes, I'm not sure how close to AGI we can get with classical computers. If quantum computers get out of the lab and into a server farm, then maybe AGI could become a thing? Or if things go proper sci-fi and someone builds a computer with lab grown brain cells. But right now, I can't see us getting creative thought out of logic gates, no matter how much money we throw at it.

 

Honestly, though, it is hard to tell what is true and what isn't with AI. One side hyping it like crazy, possibly for funding, the other determined to shit upon it, possibly as it may threaten their job.

 

 

[Biohazard777]

7 hours ago, LAwLz said:

You are missing the forest for all the trees.

 

The process involves thinking about something, evaluating the answer, rethinking it, evaluating again, and so on. This undeniably fits the definition of reasoning. It's essentially the same process you used to arrive at your answer. While you may have employed a different technique and approach, the fundamental steps remain consistent: analyze, consider an answer, verify its correctness, and if necessary, rethink and retest. That seems to fit the definition of reasoning to me.

 

It's important to recognize that something can still be valid even if it's not the most optimal approach. Your method of solving the question is, in my opinion, superior, but that doesn't negate the fact that a different approach can also constitute reasoning.

 

 

Although I am not sure how this is relevant to the topic at hand.

We are side-topic now, not off-topic but clearly not on-topic anymore, hah.
 

Sure, the approach doesn't need to be optimal, and it can have (many) failed attempts/paths along the way.
 

But I fail to see the reasoning behind the final and correct answer: "The answer to this question contains eight words".
Can you repeat what OpenAI did?
That is the difference between my and AI answers (besides efficiency, which isn't relevant to the discussion about reasoning).
 

I didn't miss the tree for the forest, thinking in an orderly and rational way IMO is not randomly picking techniques (in this case output sentences) aka throwing shit at the wall and then validating the output.

PS

Spoiler

-- correct

-- correct


-- wrong

And let's see what happens when I try correcting it with a prompt:


This is pretty much what I've seen o1 do in your example...
It didn't use reason to come to the original answer, it threw shit at the wall and then instead of waiting for user input it ran it through a loop till the output matches what the original goal was.
Again, that is not reason... Heck, there is more "reasoning" involved in the 2nd answer from 4o "Let me count..." than o1 screenshot.
Also, doesn't it remind you of that DeepMind paper you linked in this thread?